Most AI agents fail in production for a boring reason: nobody can verify what they actually did. Here’s the receipts-first rule for building agents people will trust, debug, and pay for.